Abstract: Soil moisture is one of the most influencing factors on the crop growth in agricultural production, particularly for the water management, yield estimation, drought monitoring and precision irrigation. Therefore, it is very necessary to rapidly and accurately detect the soil moisture. Fortunately, deep learning techniques have been widely applied into the soil moisture inversion in recent years. However, much efforts have been focused mainly on the optimization of model structures. It is still lacking to explore the hyperparameter settings of the models. In this study, an optimization strategy was proposed for a convolutional neural network (CNN) model using rime optimization (RIME), in order to improve the performance of soil moisture inversion in winter wheat farmlands. The polarization decomposition was also combined to correct the impact of vegetation on the accuracy of soil moisture inversion in the vegetation-covered areas The reason was that the vegetation was negatively correlated with the soil moisture inversion. The RIME was then employed to optimize the hyperparameters of CNN, in order to form the RIME-CNN model. Subsequently, the RIME-CNN model was utilized to adaptively extract the feature parameters. Soil moisture was then estimated to regularize and feed into the feature parameters using support vector regression (SVR). Additionally, the wealth polarization characteristics were contained in fully polarized data. Various techniques were employed to perform the polarization decomposition on the RADARSAT-2 data. As such, the wide ranges of characteristic parameters were acquired after polarization. The original feature space of SAR data was further enriched to eliminate the data redundancy for the convergence of the network. Mutual information (MI) was also used to optimize the characteristic parameters. Synthetic aperture radar (SAR) and optical remote sensing data were used to invalidate the efficacy of the RIME-CNN-SVR model in the soil moisture inversion of the winter wheat farmlands. The results showed that: 1) The high accuracy of soil moisture inversion was achieved to improve the correlation between surface backscatter coefficient and soil moisture. Therefore, the polarization decomposition was effectively weakened the interference of vegetation. Among them, there was the highest correlation between the surface backscatter coefficient under HH polarization and soil moisture. 2) The MI was employed to optimize the feature. The unnecessary feature parameters were effectively reduced the data redundancy. The performance of network training was enhanced for the high accuracy of soil moisture inversion. 3) The RIME-CNN-SVR model was obtained in the higher inversion accuracy, compared with the CNN, RIME-CNN and CNN-SVR models, in which the determination coefficient was 0.72, the root mean square error (RMSE) was 2.78%, and the mean absolute error was 2.2%. At the same time, the RIME-CNN-SVR model was also feasible and suitable for the inversion of soil moisture in the winter wheat fields. The finding can also provide the accurate and reliable means of soil moisture monitoring for agricultural production.